# Skin-lesion test separates diagnosis accuracy from discharge risk

> CNN led diagnosis; its preset discharge rule missed more melanomas.

By BIGCHANGE.AI Editorial

Published: 2026-10-26T00:16:18.025Z
Updated: 2026-10-26T00:16:18.025Z
Canonical: https://bigchange.ai/blog/skin-lesion-test-diagnosis-discharge-threshold

![Charcoal illustration of a dark handheld dermatoscope resting diagonally in a stainless clinical tray.](https://bigchange.ai/api/media/file/dermatoscope-tray-concept-hero-v1.png)
AI-generated conceptual illustration by BIG CHANGE.

Magnus Lynch's [October 25 test](https://lynchlab.net/llm-skin-cancer) compares diagnosis with discharge recommendations.

## The big change

- **What changed:** Six-label accuracy favored a convolutional neural network (CNN).
- **Why it matters:** Correct labels cannot establish safety of recommendations to stop assessment.
- **What to watch:** The incomplete threshold-matching rule limits discharge comparisons.

## What the numbers measure

First-choice accuracy counts correct diagnoses. A discharge label means no further assessment; no patients were discharged.

Of 200 selected, biopsied lesions (100 keep, 100 discharge; 30 melanomas), CNN labelled 125 correctly; author-labelled GPT-6 Astra and Claude Fable 5.1 scored 84 and 85. Each language model was run once per case; exact revisions and run dates are undisclosed.

At its preset threshold, CNN recommended discharge for 18/30 melanomas; language models, three and four. A threshold change improved CNN's count. BIG CHANGE did not rerun it.

[MILK10k](https://api.isic-archive.com/doi/milk10k/) pairs clinical and dermoscopic images. This cohort cannot establish routine-care performance; public release leaves training exposure unresolved.

## Sources

- [Magnus Lynch skin-cancer comparison](https://lynchlab.net/llm-skin-cancer) — Author protocol; model revisions and run dates undisclosed.
- [MILK10k, ISIC Archive](https://api.isic-archive.com/doi/milk10k/) — Dataset description, paired-image design and image licence; distinct from the author's selected sample.